CVtreeMLE: Efficient Estimation of Mixed Exposures using Data Adaptive Decision Trees and Cross-Validated Targeted...

Area of Science:

• Biostatistics
• Epidemiology
• Environmental Health Sciences

Background:

• Traditional methods for mixed exposure analysis often rely on parametric models or assess exposures independently, leading to biased estimates of joint impacts.
• Existing mixture methods like ridge/lasso regression and principal component regression have limitations including linear assumptions, user-defined interactions, and loss of interpretability.
• Advanced methods such as quantile g-computation and Bayesian kernel machine regression have biases or computational challenges, lacking robust summary statistics for dose-response relationships.

Purpose of the Study:

• To introduce a novel non-parametric statistical machine learning approach for causal inference of mixed exposures.
• To provide a robust method that identifies data-adaptively determined decision trees for mixed exposure analysis.
• To offer interpretable results and valid inference for target parameters, addressing limitations of existing statistical models.

Main Methods:

• Utilizes non-parametric decision tree methods to identify partitions in the joint-exposure space that explain outcome variance.
• Employs a novel approach to decision tree inference that avoids overfitting by not using the full dataset for both node identification and inference.
• The CVtreeMLE R package is developed to implement these state-of-the-art statistical methodologies.

Main Results:

• The CVtreeMLE package offers a flexible, non-parametric alternative to potentially biased generalized linear models (GLMs) for mixed exposure analysis.
• It enables data-adaptive determination of the best fitting decision tree for evaluating combined exposures.
• Provides interpretable results and valid statistical inference for causal effects of mixed exposures.

Conclusions:

• The CVtreeMLE R package equips researchers with advanced statistical tools for causal inference in mixed exposure settings.
• It overcomes limitations of traditional and existing advanced methods by offering a flexible, non-parametric, and interpretable approach.
• Facilitates more accurate assessment of the joint impact of multiple environmental exposures on health outcomes.